---

name: optimization-workflows description: "Run, control, and debug pymoo optimization workflows with algorithms, termination, callbacks, ask-and-tell loops, and result interpretation." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

optimization-workflows

Use this sub-skill when a task asks how to run pymoo optimizers: choosing an algorithm, calling minimize, setting termination, controlling stochastic reproducibility, using callbacks or display/output hooks, stepping an algorithm manually, adapting an ask-and-tell loop, interpreting Result, or diagnosing an empty/infeasible optimization outcome.

Route first

• Define objective/constraint functions, Problem output shapes, bounds, or built-in test problems with problem-modeling.
• Customize variables, sampling, crossover, mutation, repair, duplicate elimination, or initialization with operators-and-variables.
• Speed up expensive evaluation, use starmap/joblib/dask/ray, or inspect compiled extensions with performance-and-parallelization.
• Compute hypervolume, IGD/GD, MCDM decisions, convergence curves, reference directions, or plots with analysis-and-visualization.
• Stay here for optimization execution control, algorithm portfolio routing, termination decisions, callbacks/display/history, and Result fields.

Fast operating checklist

1. Normalize the optimization contract: pymoo minimizes every objective. Confirm the problem exposes n_obj, n_ieq_constr/n_eq_constr, bounds, and finite F/G/H values before tuning algorithms.
2. Choose an algorithm by objective count and variable type: use single- objective algorithms such as GA, DE, PSO, CMAES, or local searches for one objective; use NSGA2, SPEA2, SMSEMOA, GDE3, or reference-direction algorithms such as NSGA3, UNSGA3, RVEA, MOEAD for multi/many-objective runs.
3. Start with functional minimize: pass (problem, algorithm, termination) plus seed and verbose=False. Remember minimize deep-copies the algorithm by default; inspect the executed copy through res.algorithm.
4. Select an explicit termination: use a tuple such as ("n_gen", 50), ("n_evals", 5000), ("time", "00:05:00"), or a termination object when defaults are too vague for verification.
5. Control stochasticity: pass seed=... and avoid hidden random draws in objective code. Compare algorithms with the same problem, seed policy, and evaluation budget.
6. Use callbacks/history intentionally: callbacks are lightweight for logging; save_history=True stores algorithm snapshots and can use substantial memory.
7. Interpret Result defensively: check res.X, res.F, res.G, res.CV, res.opt, res.algorithm.evaluator.n_eval, res.exec_time, and whether the best solution is feasible before reporting success.

Open the bundled references

• API reference: signatures and field contracts for minimize, algorithms, termination, callbacks, display/output, and Result.
• Algorithm selection: representative SOO, MOO, many-objective, dynamic, and preference-guided algorithm routes with import paths and fit signals.
• Workflows: functional minimize, direct algorithm stepping, ask-and-tell, external evaluation, checkpoint/debug, and comparison patterns.
• Troubleshooting: termination lookup failures, infeasible/empty results, stochastic drift, callback/history mistakes, and common API misuse.

Bundled script

• scripts/run_minimize_smoke.py: safe CPU-only quickstart smoke that runs a tiny NSGA-II/ZDT1 optimization and asserts the result shape and evaluation count.